AI4EOSC Platform
The AI4EOSC Platform is a software ecosystem built to support the development and deployment of AI, ML, and DL models and applications within the European Open Science Cloud (EOSC). It comprises a range of high-level services, with a strong focus on the Software as a Service (SaaS) layer, built on modern technologies to ensure smooth integration and efficient performance. Through its commitment to open standards, open APIs, and reproducibility, the AI4EOSC Platform helps researchers and data scientists accelerate innovation and drive AI-enabled scientific progress within the EOSC framework.
AI4OS (AI for Open Science)
AI4OS is a collection of software and tools for building cloud platforms that support the development, training, sharing, and deployment of AI applications using distributed cloud resources. It was developed as part of the AI4EOSC project, and it is being used to build AI platforms for a variety of use cases.
EOSC-ARENA will follow an incremental development methodology,
following agile principles. For more information on the technical Work Packages see also the project plan.
Implementation will span two complementary activities: infrastructure and service integration for GenAI (i.e. a technological infrastructure and platform) and Agentic GenAI technologies for science and research (i.e. the development and implementation of GenAI assets and an agentic AI system).
EOSC-ARENA will embrace distributed and federated cloud infrastructures together with HPC resources, offering a comprehensive set of services taking as a starting point the AI4EOSC platform and its architecture.
EOSC-ARENA extends the platform beyond model training, inference and fine-tuning to also orchestrate agent executions, with secure, pluggable agent execution, agent observability and an extended model marketplace.
A training layer to deliver a federated training and fine-tuning set of components following industry standards (Flower, NVIDIA-FLARE), providing multimodal capabilities and efficient data access.
Serving layer based on efficient model access through vLLM, extending it with federated capabilities (to be able to exploit the EOSC federated nature), semantic routing, model context switching, and API key management.
A serving layer for non-LLM based models, modeled on a microservices architecture and adapted to meet the EOSC’s requirements for quality assurance, provenance tracking, and adherence to FAIR principles.
Training, inference and knowledge retrieval platform that will execute on federated distributed computing platforms based on Cloud technologies (e.g. OpenStack, but also with the capabilities to span to commercial providers) on which Container Orchestration Frameworks (i.e.. Nomad/Kubernetes) customized with the EOSC-ARENA platform will be dynamically deployed from standards-based descriptions.
HPC used for federated model training. These will be made available via open-source licenses for reproducibility and ease of deployment on European distributed infrastructures such as EGI and the EOSC EU Node, one of the reference EOSC Nodes.